10 Sep NYU Study: AI Model Using Longitudinal 3D Mammography Outperforms Standard Risk Tools for 5-Year Breast Cancer Prediction
MedicalResearch.com Interview with:
Dr. Shen Yiqiu
Yiqiu Shen, PhD
Assistant Professor, Department of Radiology
NYU Grossman School of Medicine
Dr. Yanqi Xu
Yanqi Xu, PhD
NYU Center for Data Science
MedicalResearch.com: What is the background for this study? What images does the model utilize?
Does the model require different and/or more imaging?
Dr. Shen and Dr. Xu: I would like to first introduce the task that this AI model is built for. Breast cancer screening is designed to detect cancers that are already present. Radiologists are trained for this task. However, breast cancer risk prediction aims to identify currently cancer-negative women who are more likely to develop breast cancer in the future. This AI model is designed to predict 5-year breast cancer risk. According to the ACR (American College of Radiology), risk-stratified screening approaches are increasingly recognized as a way to personalize breast cancer screening — yet most existing risk models rely on broad clinical factors rather than the detailed imaging information available in modern mammography systems.
No radiologists or clinicians are trained for the risk prediction task. In the current practice, breast cancer risk is estimated by some existing risk calculators, such as the Gail Model and the Tyrer-Cuzick model. Most existing clinical risk models rely largely on factors such as age, family history, reproductive history, and breast density. These factors are overly broad and not personalized to individual patients. As a result, the accuracy of these risk models is limited. More recently, artificial intelligence has shown that mammographic images themselves contain additional information about future breast cancer risk.
Most previous imaging-based AI risk models have been developed using conventional two-dimensional mammography. However, digital breast tomosynthesis, or DBT, has become a predominant form of breast cancer screening in the United States and provides a more detailed, quasi-three-dimensional representation of breast tissue. We therefore asked whether AI could use not only the 3D information contained in DBT, but also changes across a woman’s prior DBT examinations, to better predict future breast cancer risk.
Importantly, the model does not require a new imaging test or additional radiation exposure. It uses DBT images that are already acquired as part of routine breast imaging, together with the patient’s age and breast density. For each examination, the model can also incorporate prior DBT examinations when available — up to 10 previous examinations in our study — allowing it to evaluate how the breast changes over time.
MedicalResearch.com: What are the main findings?
Dr. Shen and Dr. Xu: We studied more than 313,000 DBT examinations from approximately 161,000 women and evaluated the model on a hold-out test cohort. The longitudinal DBT model achieved an AUC of 0.721 for 5-year breast cancer risk prediction, compared with 0.707 when only the current DBT examination was used. This suggests that prior examinations provide additional information about future risk.
The model also performed better than two established approaches. It outperformed Mirai, an AI risk model based on conventional 2D mammography, with 5-year AUCs of 0.721 versus 0.687. In a separate matched cohort, our model also outperformed the widely used Tyrer-Cuzick clinical risk model, with AUCs of 0.676 versus 0.563.
One particularly interesting finding was that the AI model could refine risk beyond breast density alone. Traditionally, increased breast density is believed to be associated with increased breast cancer risk. However, it is a relatively coarse-grained and subjective rating given by radiologists. Our study suggested that AI models might be able to investigate breast tissue patterns that can provide better risk stratification than using density alone. For example, among women with extremely dense breasts (highest breast density), approximately 40% were classified by the model as average risk, and this group had an observed 5-year cancer incidence of only 0.8%. Conversely, among women with fatty breasts (lowest breast density), which are generally considered lower risk, the model identified approximately 15% as high risk, with an observed cancer incidence of 2.6%. This illustrates that women with the same breast-density category can have substantially different underlying risks.
MedicalResearch.com: How accessible would this tool be for wider dissemination?
Dr. Shen and Dr. Xu: One attractive aspect of this approach is that it could potentially be integrated into existing breast screening workflows because it uses imaging that is already routinely collected. In principle, a woman could undergo her regular DBT examination and have an individualized risk estimate generated from the current and prior examinations without undergoing an additional test or providing an extensive new set of clinical information.
From my perspective, the current study should be viewed as a development and internal validation study rather than a tool that is ready for widespread clinical use. The model was developed retrospectively at a single academic health system, and the DBT examinations were acquired using systems from a single vendor. Before clinical deployment, we need to demonstrate that the model generalizes across institutions, patient populations, imaging equipment, and clinical workflows. Prospective evaluation will also be important to determine whether using these predictions actually improves screening decisions and patient outcomes.
MedicalResearch.com: What recommendations do you have for future research as a result of this study?
Dr. Shen and Dr. Xu: There are several important next steps. First, the model should be externally validated in multi-institutional and multivendor datasets, including larger and more diverse patient populations. This is particularly important because we observed some variation in performance across racial and ethnic subgroups. Second, we would like to combine the imaging information captured by the AI model with other established risk factors, such as family history, reproductive history, prior breast procedures, and genetic information. Imaging and clinical information likely capture complementary aspects of risk, and integrating them may provide a more complete individualized assessment. Third, future models should provide greater interpretability so that we can understand which breast imaging patterns — and which changes over time — are associated with elevated or reduced risk. Findings in this direction could potentially uncover new knowledge underlying cancer carcinogenesis and help us understand breast cancer better. Finally, the most important step will be prospective clinical studies evaluating whether AI-based risk prediction can safely guide decisions such as screening frequency, supplemental MRI or ultrasound, and preventive interventions.
MedicalResearch.com: Is there anything else you would like to add?
Dr. Shen and Dr. Xu: I think an important broader message from this study is that a screening image may contain considerably more information than simply whether cancer is visible today. The same examination may also provide a window into a patient’s future risk. By analyzing subtle characteristics of breast tissue and how they change over time, AI may eventually allow routine screening examinations to function as dynamic biomarkers of future breast cancer risk. Our goal is not simply to identify more women as being at high risk. Ideally, better risk prediction would allow us to identify both women who may benefit from more intensive screening or prevention and women who may safely avoid unnecessary additional testing. Our findings are encouraging in this regard, but prospective clinical validation will be essential before these predictions are used to change patient care.
Disclosures: None disclosed.
Citation:
Xu Y, Heacock L, Park J, Pasadyn FL, Lei Q, Lewin A, Geras K, Moy L, Schnabel F, Shen Y. Predicting 5-Year Breast Cancer Risk From Longitudinal Digital Breast Tomosynthesis: A Single-Center Retrospective Study. American Journal of Roentgenology. Published online August 12, 2026. doi:10.2214/AJR.26.34951
For a broader overview of how AI is being applied to breast cancer screening and what the evidence shows about mammography-based risk prediction models, see this MedicalResearch.com overview of AI in breast cancer screening — what the evidence shows.
Disclaimer: The information on MedicalResearch.com is provided for educational purposes only, and is in no way intended to diagnose, cure, or treat any medical or other condition. Some links are sponsored. MedicalResearch.com and Eminent Domains Inc. do not warrant or endorse products or claims made by third party links. Always seek the advice of your physician or other qualified health provider and ask your doctor any questions you may have regarding a medical condition. In addition to all other limitations and disclaimers in this agreement, service provider and its third party providers disclaim any liability or loss in connection with the content provided on this website.
Last Updated on September 10, 2026 by Marie Benz MD FAAD